EP1368714B1 - Estimation method for a variable of a process of process manufacturing using a reference vector method - Google Patents

Abstract

The invention relates to a method according to which a process variable of a process (1) of process manufacturing is estimated during the progression of the process (1) by means of a reference vector model (2). For preliminarily training said model, a number (N) of pairs of values is defined that comprise one vector of detectable signals (x1<i>, , xn<i>) of the process (1) and one measuring variable (y<i>) each for the process variable to be estimated that corresponds to said measured value. At least one further pair of values that is detected by means of the process (1) is defined for online post-training.

Description

• wobei eine Anzahl von Wertepaaren vorgegeben wird,
• wobei jedes Wertepaar einen Vektor von erfassbaren Signalen des Prozesses und einen mit diesem Vektor korrespondierenden Messwert für die zu schätzende Prozessgröße aufweist,
• wobei anhand der Anzahl von Wertepaaren ein empirisches Modell für die Abhängigkeit der Prozessgröße von den erfassbaren Signalen erstellt wird,
• wobei aufgrund des Modells während des Ablaufs des Prozesses anhand mindestens eines weiteren Vektors von erfassbaren Signalen ein Schätzwert für die Prozessgröße ermittelt wird,
• wobei dem Modell mindestens ein weiteres Wertepaar vorgegeben wird, dessen Vektor während des Prozesses ermittelt worden ist und dessen Messwert nach dem Ermitteln des Vektors ermittelt worden ist,
• wobei das erstellte Modell anhand des mindestens einen weiteren Wertepaares auf einem Prozessautomatisierungsrechner angepasst wird.

The present invention relates to an estimation method for a process variable of a process in the basic material industry,

• where a number of pairs of values is specified,
• wherein each pair of values has a vector of detectable signals of the process and a measured value corresponding to this vector for the process variable to be estimated,
• an empirical model for the dependence of the process variable on the detectable signals is created on the basis of the number of value pairs,
• an estimated value for the process variable being determined on the basis of the model during the course of the process using at least one further vector of detectable signals,
• the model being given at least one further pair of values, the vector of which was determined during the process and the measured value of which was determined after the vector was ascertained,
• wherein the model created is adapted on the basis of the at least one further pair of values on a process automation computer.

Such estimation methods are e.g. based on neuronal Networks, generally known. An example is the DE 197 27 821 A1.

The estimation methods previously used in the raw materials industry like neural networks or classic regression methods have problems. The training effort often scales unfavorable for the model with the dimension of the input vectors. In the worst case, the training effort increases exponentially on. The training process is often based on one nonlinear optimization method. This can cause that the global optimum may not be found and / or the computing time requirement becomes too large.

These problems come into play even more when the model training not offline, but online, i.e. on one of the process automation computers (process control or control) of the system. Because for training (i.e. creating or adapting the empirical model based on the value pairs) is online - in contrast to offline - usually only a relatively short and clearly limited Period of time available. Furthermore, the training process run absolutely reliably online.

The object of the present invention is a To create an estimation method for an empirical model that reliable even with a higher-dimensional input variable vector and can be trained with acceptable computing effort.

The problem is solved in that the model is a support vector model is.

Because with support vector methods, the effort for training increases of the model both in terms of the number of people to train used input / output value pairs as well as regarding the dimension of the input quantity vector only linear. suitable Training procedures for the model are "denomination methods" (English "chunking methods"), e.g. those according to Joachims or the "Sequential Minimal Optimization" (English "Sequential Minimal Optimization ") according to Platt. Because these processes work reliable and fast.

a) Ein Stützvektormodell ist (bis auf einen Offset) eine Linearkombination von symmetrischen Kernfunktionen, die sich mathematisch als Skalarprodukt zweier identischer Funktionen schreiben lassen.

b) Die Koeffizienten in der in (a) genannten Linearkombination werden durch Lösen eines konvexen Optimierungsproblems (z.B. eines linearen oder quadratischen Programmierungsproblems) bestimmt.

Support vector methods or support vector models are characterized by the following properties:

a) A support vector model is (apart from an offset) a linear combination of symmetric core functions that can be mathematically written as a scalar product of two identical functions.

b) The coefficients in the linear combination mentioned in (a) are determined by solving a convex optimization problem (eg a linear or quadratic programming problem).

Support vector methods and support vector models are e.g. Tie monographs

"Advances in Kernel Methods - Support Vector Learning" by B. Schölkopf, C.J.C. Burges, A.J. Smola, MIT Press, Cambridge (Mass.), London 1999 and

"An introduction to Support Vector Machines and other kernel-based learning methods "by N. Cristianini, J. Shawe-Taylor, Cambridge University Press 2000

described in detail. It also includes the above Training procedures according to Joachim and Platt are shown and called further specialist literature.

In empirical modeling and especially in support vector methods a distinction is made between models in which the output size depends on the input quantity vector, and models, where the output quantity is only discrete values, in particular can assume binary values (i.e. 0 and 1). The former Models are also used in the literature as regression methods referred to, the second mentioned as classification or pattern recognition process.

Regression methods are usually used to predict physical process variables within process automation (Process control and regulation) of a plant, during classification or pattern recognition processes e.g. serve to detect faults in the system.

If the model is analytically partially based on the detectable signals is derivable, the online computing effort is minimal.

The symmetric core functions can e.g. B. Gaussian radial Basic functions.

If the at least one other pair of values is cached and adapting the model based on the minimum another pair of values takes place at a later point in time, the model can be adjusted during times in which the computing power of the used to adapt Calculator is not otherwise needed.

If the dependence of several process variables on one vector of detectable signals described by an empirical model such a model is equivalent to a corresponding number of empirical models that describe the dependency of one of these process variables.

In principle, the process can be of any nature. In particular can it be a coal and ore processing process, a smelting process, a steelmaking process, a casting process, a rolling process and a pulp and paper production or processing process.

FIG 1

eine Anlage der Grundstoffindustrie und

FIG 2

ein Ablaufdiagramm.

The following description of an exemplary embodiment describes a method for creating online-capable empirical models using support vector methods. Show in principle

FIG. 1

a plant of the basic material industry and

FIG 2

a flow chart.

1 for a process 1 of the raw materials industry a process size of process 1 can be estimated. The process 1 can be any process 1 in the raw materials industry his. A smelting process is shown schematically in FIG 1a, a steel making process 1b, a casting process 1c and a rolling process 1d. But there are also others Processes in question, especially a preparation process for Coal or ore and a production or processing process for pulp or paper.

A certain is first used to estimate the process size online Kind of a support vector model 2 together with one Training method selected. It should be noted that the training method in terms of computing time and reliability is online-capable, i.e. on a process automation computer 2 'can be used for process 1. Because on the process automation computer 2 'of the system is the estimation method preferably implemented. As a support vector model 2 can e.g. - but not necessarily - that in US-A-5,950, 146 described support vector model according to Vapnik become.

Furthermore, a number N of value pairs is collected from the installation on which process 1 is running (or a similar installation). The value pairs each consist of a vector of detectable signals x ₁ ⁱ , ..., x _n ⁱ , on which the process variable to be estimated depends, and a measured value y ⁱ corresponding to this vector for the process variable to be estimated (i = 1,. .., N). This number N of pairs of values is used for the offline training of a support vector model 2 of the selected type. An empirical support vector model 2 for the dependence of the process variable on the detectable signals is therefore created on the basis of the number N of value pairs. The pre-trained support vector model 2 is then implemented together with the selected online training method on the process automation computer 2 'of the system.

For online estimation of the process variable, a vector of detectable signals x ₁ ,..., X _n of process 1 is fed to the support vector model 2 as an input variable vector. The support vector model 2 then uses the supplied vector to determine an estimated value y * for the process variable during the course of the process 1.

The actual value y of the process variable can often also be determined using process 1 (but always only afterwards). This value y is fed together with the input variable vector to the support vector model 2 for online training. Using the value pair formed from the input variable vector and the process variable, i.e. on the one hand the vector of detectable signals x ₁ , ..., x _n and on the other hand the corresponding measured value y for the process variable, the support vector model 2 is then retrained online (ie on the process automation computer 2 ') or adjusted.

Model 2 is preferably created and adapted according to a "chunking method" like her z. B. is described by Joachims or as they are called "Sequential Minimal Optimization" ("Sequential Minimal Optimization ") has been developed by Platt.

With sufficient free computing capacity of the process automation computer 2 'can adapt the model 2 immediately after determining the new pair of values. In the As a rule, however, it will make more sense to add the if necessary, several further pairs of values - to buffer and that Customize model 2 later with the up to then accumulated pairs of values.

The process above on the process automation computer 2 'is explained again in connection with FIG 2.

According to FIG. 2, the model 2 is given a number N of value pairs in a step 3. Each pair of values consists of a vector of detectable signals x ₁ ⁱ , ..., x _n ⁱ and a measured value y ⁱ corresponding to this vector for the process variable to be estimated. Based on these pairs of values, the model 2 is then trained in a step 4.

A vector of detectable signals x ₁ ,..., _{Xn is then} input to the support vector model 2 in a step 5. In a step 6, it then determines an estimated value y * for the process variable. The actual value y of the process variable is then determined in a step 7. The vector of detectable signals x ₁ , ..., x _n is buffered together with the value y in a step 8. In a step 9, a query is then made as to whether model 2 should be retrained. Depending on the result of the query, the system either jumps back to step 5 or does a subsequent training of model 2 in step 10.

k ist eine symmetrische Kernfunktion, die sich mathematisch als Skalarprodukt zweier identischer Funktionen schreiben lässt. Als Kernfunktionen k kommen z. B. Gaußsche radiale Basisfunktionen k(xⁱ, x^j) = exp(-|xⁱ-x^j|²/c) in Frage. Alternativen, die anwendungsabhängig evtl. günstiger sein können, sind in den beiden genannten Monographien angegeben.Support vector models for modeling continuous process variables are e.g. B. from the shape

k is a symmetric core function that can be mathematically written as a scalar product of two identical functions. As core functions k come z. B. Gaussian radial basis functions k (x ⁱ , x ^j ) = exp (- | x ⁱ -x ^j | ² / c) in question. Alternatives that may be cheaper depending on the application are given in the two monographs mentioned.

The entire support vector model 2 is a linear combination of these functions k. x ^j (j = 1, ..., N) are the input vectors of the value pairs used for training. w _j and b are coefficients to be determined by the training process.

The core functions k are analytically given functions, which are generally continuously differentiable. It is therefore possible to form the partial derivatives according to all components of the input quantity vector x in an analytical form. Because the entire support vector model 2 is a linear combination of these core functions k, so-called model sensitivities (= partial derivations of the process variable to be estimated or of model 2 according to the detectable signals x ₁ , ..., x _n ) can also be analyzed beforehand form and store in the support vector model 2. Compared to numerically determined sensitivities, model sensitivities stored analytically have the advantage of higher accuracy with generally less computation effort.

If not a continuous, but a discrete process variable to be modeled is described in the above A constant intermediate quantity is first determined, to which a suitable discretization function, e.g. the signum function is applied.

Above was an estimation procedure for a single scalar Process size described. If the dependence of several Process variables from a vector of detectable signals an empirical model 2 is to be described, then is a such model 2 equivalent to a corresponding number of empirical models that dependence one of these Describe process variables.

Claims (10)

1. Estimation method for a process variable in a process (1) in the process manufacturing industry,

   with a number (N) of value pairs being predetermined,

   with each value pair having a vector of detectable signals (x₁ ⁱ, ..., x_n ^I) for the process (1) and having a measured value (yⁱ) which corresponds to this vector, for the process variable to be estimated,

   with an empirical model (2) for the relationship between the process variable and the detectable signals (x₁, ..., x_n) being produced on the basis of the number (N) of value pairs,

   with an estimated value (y*) for the process variable being determined on the basis of the model (2) while the process (1) is being carried out, on the basis of at least one further vector of detectable signals (x₁, ..., x_n),

   with at least one further value pair being predetermined for the model (2), whose vector has been determined during the process (1) and whose measured value (y) has been determined after the determination of the vector,

   with the model (2) that is produced being matched to a process automation computer on the basis of the at least one further value pair, and

   with the model (2) being a reference vector model (2).
2. Estimation method according to Claim 1, characterized in that the process variable can assume any discrete values.
3. Estimation method according to Claim 1, characterized in that the process variable can assume continuous values.
4. Estimation method according to Claim 3, characterized in that the model (2) can be derived analytically, partially on the basis of the detectable signals (x₁, ..., x_n) .
5. Estimation method according to Claim 3 or 4, characterized in that the reference vector model (2) contains Gaussian radial basic functions as symmetrical core functions (k).
6. Estimation method according to one of the above claims, characterized in that the production and matching of the model (2) are carried out using a chunking method.
7. Estimation method according to Claim 6, characterized in that the chunking method is in the form of sequential minimal optimization.
8. Estimation method according to one of the above claims, characterized in that the at least one further value pair is temporarily stored, and in that the matching of the model (2) is carried out on the basis of the at least one further value pair at a later time.
9. Estimation method for two or more process variables in a process in the process manufacturing industry, characterized in that an estimation method according to one of the above claims is used for each of the process variables.
10. Estimation method according to one of the above claims, characterized in that the process (1) is a carbon and ore conditioning process, a smelting process (1a), a steel production process (1b), a casting process (1c), a rolling process (1d) or a pulp and paper production or processing process.

Images